Do we need polls? Why Twitter will not replace opinion surveys, but can complement them. Javier Sajuria Department of Political Science University College London Jorge Fábrega School of Government Universidad Adolfo Ibáñez Can we observe public attitudes using social media data? Or more concretely, can we disregard polls as the preferred method for observing public opinion? Our comparison between public opinion survey data and Twitter data from the 2013 Chilean presidential election shows a nuanced picture. We use social network analysis to estimate political positions of Twitter users and estimate their sentiment towards public issues based on their public tweets. We work with the public opinion data from the Chilean CEP survey to compare the Twitter data to the polls. Our focus is on the relationship between political position — proxied by their support for a candidate — and their views about given relevant political topics. The results show that, in most cases, support for certain policies has a correlate in the online world, with a positive tone of tweets. However, there are some interesting differences among supporters of various candidates. Those who support the leading candidate tend to tweet with a more positive tone, regardless of the issue. On the other hand, supporters of other candidates are less likely to tweet with a positive tone, even when it is about a topic they support. These findings show that Twitter data can provide interesting insight about how people frame political discussions, depending on the electoral viability of the candidate they support. Keywords— political position estimation, social media, network homophily, electoral viability Introduction Monitoring and using social media to understand — or influence — public opinion is not a new thing. Companies, political parties and organisations alike are keen to observe what their followers say, what people are commenting on their Facebook pages, and what is said in the comments sections of Youtube and Instagram. Moreover, a great deal of work has been done in building social media teams in charge of both engaging and analysing what people exchange through these platforms. To some extent, these phenomena have questioned whether traditional, more expensive, ways to observe public opinion are still required. The regular route for understanding public opinion, both at the consumer and the political levels, relies heavily on surveys. These instruments present their own advantages depending on the scope of the research. Moreover, they enjoy a fair amount of validity among the scientific community as proper instruments to analyse public attitudes. Twitter, on the other hand, has been widely contested by the academic community as a valid way to analyse public attitudes and behaviour. Different attempts to predict election outcomes from Twitter have failed, and scholars (Gayo-Avello 2012; DiGrazia et al. 2013) have argued about the usefulness of social media data to understand large-scale political events. The same has been argued in relation to other events, such as the Eurovision contest or popular TV Shows (e.g. The X Factor show). The underlying consensus is that Twitter does not present the conditions required by traditional research approaches to produce accurate forecasts. Hence, some recent attempts have pursued a different route: comparing Twitter data to opinion polls. Some recent efforts (e.g. Beauchamp 2013) aim to forecast candidates’ approval ratings by matching them with Twitter data from the previous period. In that way, the goal is no longer in predicting elections (that is left to opinion polls), but to analyse how close are the discussions on Twitter to more valid representations of public opinion. This chapter aims to expand this line of research by using two different strategies. On the one hand, we use retweet networks to estimate the political position of Twitter users. Second, we take the content of the tweets from those users to compare their views on different topics with data from public opinion surveys. Scholars (Ansolabehere, Rodden, and Snyder 2008; Bartels 2010; Iyengar, Sood, and Lelkes 2012) have already established the presence of a relationship between political positions and attitudes towards public issues, such as equal marriage and the electoral system. We use the estimation from Twitter data to compare the results with opinion polls and provide a more informed picture of when, if possible, social media can substitute or complement them. We use data from Chile for 2013, focusing in the period before the presidential election. The Twitter data was gathered from 17 September to 17th December of 2013 using the Twitter public streaming API. Survey data, on the other hand, comes from a mainstream source in Chile: The seasonal survey from the Centro de Estudios Públicos (CEP). We discuss the validity issues of each source and the strategy to assess their accuracy. Our results show that, according to the expectations, Twitter data is still not appropriate for substituting opinion polls. However, there is an interesting story to be told in relation to candidate support and tone. In the Chilean case, the supporters of the leading candidate, and later President, Michelle Bachelet, are more likely to express their views on Twitter with a positive tone. Moreover, when the support for Bachelet predicts significantly the support for certain policies (according to survey data), there is a correspondence in the positive tone of the tweets from her Twitter supporters. This is something that does not happen in the case of those who support other candidates. In other cases, we see supporters that tend to be less likely to use a positive tone on their tweets, even when survey data says that, on average, they support the policies that are talking about. This is consistent on our hypothesis that electoral viability is related to higher likelihood of a positive tone in the tweets. This chapter will go as follows. First, we discuss the literature on opinion formation and the role of political position. Then, we move into the discussion of using social media data to forecast political events and understand public attitudes. This is followed by a discussion of the Chilean case and the elements of electoral viability. We then explain the methods used to estimate political positions and filter the relevant topics. The results are presented to demonstrate how we derive the conclusions stated above. In our discussion, we extrapolate from these results to make a compelling case of how much researchers should rely on these sources and what is the actual potential of new media for valid academic research. Literature Review Ideology and Public Opinion Political ideology is relevant for public opinion. This is a bold statement, but not unjustified. Zaller argued in 1992 in his well-known RAS model that people who are more politically aware tend to have more stable and defined attitudes. For Zaller, ideology was a product of this awareness. The more aware a person is, the more stable are their ideological positions. Then, people with more defined ideology or systems of belief will look for information from partisan voices. That is, liberals will search for opinions from liberal elites, and will reinforce their own liberal views. This will reflect on preferences for public policies (such as redistribution or welfare in the case of liberals), and approval ratings. Conversely, Zaller claims that less aware people have also less stable ideological positions. In turn, this will reduce their “attitude constraints” creating inconsistency. In short, more ideology leads to more consistent attitudes. The empirical evidence supports this view. For example, Bartels (2005) studies how views on tax reform in the US are explained by ideology and levels of education. Converse (1975) makes a similar case in relation with voting behaviour, while Dalton (2000) analyses the role of party identification (usually used as a measure for ideological position) in today’s politics. Outside the US, there is growing body of literature on the topic. López-Sáez & Martínez-Rubio (2005) explain how ideological positions change the level of credibility in governmental information. Based on the case of the 11-M terrorist attacks in Madrid, they found that right-wing people believed that their voting behaviour had been affected more by official information, while left-wing respondents were more influenced by unofficial information. Estimating people’s political position is, then, extremely relevant to understand political attitudes. Traditional measures have relied on survey questions where people can position themselves in 1-10 (or 1-7) scales, indicate their preferences for the existing political parties, or self-identify as liberals or conservatives. The validity of each of these measures has been widely discussed (see discussion on Ansolabehere and Hersh 2012) and is usually dependent on the political system of each country. In multiparty systems, measures of left-right scale might be an over-reduction of the complexity in which people can position politically. Accordingly, countries with two- party systems are more suitable for such scales. In the case of Chile — the case under study — the presence of two big coalitions for the last 20 years allows us to use methods that are similar to two-party systems. Using Social media for forecasting elections and understanding public attitudes Since Nate Silver’s fairly accurate predictions of the last US elections, forecasting events has become an attractive topic. Nowadays, we can find statistical models to predict the outcome of the Fifa World Cup, the winner of the Eurovision contest, or the next armed conflict in the world. With unequal results, the advancements of forecasting models rely heavily on the quality of the information they use to base their predictions. For examples, attempts to produce similar election forecasts in Chile (Bunker n.d.) have failed, mainly due to the inability of pollsters to estimate turnout. Similar situations can be observed in other Latin American countries, where the low quality of survey data produces bad forecasts.